Know · the meeting, as one thing

A meeting was always four things.
Now it is one.

The invitation, the people, the hour and the recording used to be four unrelated records — a transcript attached to a note, and a calendar event that existed only for as long as the fetch that returned it. LumenQube now keeps the meeting itself, and hangs the rest off it.

One object, joined across two sources

The hard part was never the transcript. It was that the appointment and the recording of the appointment had no way of knowing they were the same event.

What the meeting holds

  • The same identity from both directions. A note linked to an event resolves to the same meeting as the event itself, so the appointment and the recording of it are one thing rather than two that look alike.
  • Who was actually there. Both calendars always sent the guest list; both threw it away and kept the count. The identities come through now.
  • The transcript, with turns and timings — and the decisions and action items drawn out of it, each carrying a quote.
  • It is registered as sensitive, so it stays out of the plaintext index along with your notes and tasks.

Carried carefully

  • Provider‑free by test, not by intention. The check scans every string value on the object for the provider’s own identifiers rather than checking a list of field names — a list is a thing you forget to update.
  • A 400‑person all‑hands truncates its list and never loses its number. Attendees are capped at 32 on the wire and again on arrival, with the count kept separately.
  • The two calendars disagree about RSVP in six different spellings. They collapse to four, once, in one place.

A citation you can check, or no citation at all

Most summarisers report a citation: the model says where it got something, and you take that on faith. This one verifies it. The difference only shows up when the model is wrong, which is exactly when it matters.

  1. The model reads the transcript and proposes a decision

    Along with the words it claims were said — an actual quote, not a paraphrase and not a timestamp.

  2. The quote is looked up LumenQube searches the transcript for those words

    Not the model’s account of where they were. The transcript is searched for the quote, and the time is read off wherever it actually landed.

  3. Unfound means dropped A claim whose quote is not there does not reach you

    It is discarded and counted — not shown with a warning, not shown with a lower confidence score. Measured against a set with three fabrications planted in it: none survived.

  4. What is left is shown with the moment it came from

    So a decision you do not remember agreeing to can be checked against the recording in one click, rather than argued about.

The honest cost, because this trade is a real one: a model that paraphrases its quote instead of copying it loses a true claim the same way it loses a false one. Recall is traded for the guarantee. The result is a shorter list than a summariser would give you, and every line on it can be checked — which is the trade worth making for a decision log and would be the wrong one for a first draft.

Where it appears, and what it will not touch

A write‑up that rearranges your own notes is not a feature.

Where it goes

  • Under the recordings strip on the note — never inside the body of the page you wrote.
  • Your page is byte‑identical after extraction. That is asserted by a test, not just intended.
  • Collapsed by default. The first version stood about 394px tall and pushed your own writing below the fold, which is the sort of thing that is invisible in source and obvious the moment you look at it.
  • Action items become real tasks rather than a bulleted list that agrees with nothing.

What is not built

  • No speaker identification. This does not diarise. Where the source labels speakers they are kept; where it does not, a turn says “speaker not identified” rather than guessing a name.
  • No person page. “Who was in that meeting” is answerable, because the attendance links exist — but a surface showing one person’s meetings, commitments and topics is not built, and nothing here implies it is.
  • Nothing runs while the laptop is shut. Meeting artifacts are gathered by this app, on this machine, when it is awake.
  • No Zoom or Teams. The calendars are Google and Outlook.

Questions

Does LumenQube join my meetings as a bot?

No. There is no participant that dials in, and nothing appears in the attendee list on your colleagues’ screens. It works from the recording and the calendar entry you already have — which also means it captures nothing you did not capture yourself.

How is a cited decision different from a summary?

A summary asserts. A cited decision carries the words that were actually said, and those words are looked up in the transcript before you see the claim — if they are not found, the claim is dropped rather than softened. The list is shorter than a summary and every line on it can be checked against the recording.

Does it know who said what?

Only where the recording already carries speaker labels. LumenQube does not separate voices, so an undiarised recording produces turns marked “speaker not identified”. Guessing would be the more impressive behaviour and the less trustworthy one.

Is the transcript uploaded?

Transcribing a saved recording sends that audio to a managed service and needs a paid plan; the live transcript uses your operating system’s own speech engine. Extraction afterwards is an AI request like any other, and what it sends is visible before it goes.

What happens to the action items?

They become tasks in the same canonical store as the rest of your work, linked back to the meeting they came from, so they appear in Today alongside everything else rather than in a separate list that quietly disagrees.

What the meeting connects to

Notes — where the transcript livesCalendar — excluded means never askedMemory & context — and how to erase itToday — what the day needs

Download freeSee LumenNotes